Radiomics for Prediction of Radiation-induced Lung Injury After Robotic Stereotactic Body Radiotherapy of Lung Cancer: Results From Two Independent Institutions


 Objectives: To generate and validate a state-of-the-art radiomics model for prediction of radiation-induced lung injury and oncologic outcome in non-small cell lung cancer (NSCLC) patients treated with robotic stereotactic body radiation therapy (SBRT).Methods: A radiomics model was generated from the planning CT images of 110 patients with primary, inoperable stage I/IIa NSCLC who were treated with robotic SBRT using a risk-adapted fractionation scheme at the University Hospital Cologne (training cohort). In total, 851 radiomic features fulfilling the standards of the Image Biomarker Standardization Initiative (IBSI) were extracted from the outlined gross tumor volume (GTV) and used to build a model for prediction of local control (LC), disease-free survival (DFS), overall survival (OS) and development of local lung fibrosis (LF) by means of a gradient-boosted ensemble of regression trees. In addition, predictive clinical and dosimetric parameters were identified from a standard univariate Cox regression analysis. The radiomics model was validated in a comparable cohort of 71 patients treated by robotic SBRT at the Radiosurgery Center in Northern Germany (test cohort).Results: Oncologic outcome did not differ between the two cohorts (OS at 36 months 56% vs. 43%, p=0.065; median DFS 25 months vs. 23 months, p=0.43; LC at 36 months 90% vs. 93%, p=0.197). Local lung fibrosis developed in 33% vs. 35% of the patients (p=0.75), all events were observed within 36 months. In the training cohort, the radiomics model was able to distinguish low-risk from high risk patients for OS, DFS, LC and LF with a high accuracy (p < 0.001). In the test cohort, the model for development of lung fibrosis retained its predictive power and could differentiate patients with a high risk for developing LF from those with a low risk (p=0.016). In contrast, the radiomics model failed to predict OS, DFS and LC in the test cohort. Also, none of the clinical and dosimetric parameters predictive for development of LF in the training cohort (GTV-Dmean, GTV-Dmax, PTV-D95%, Lung-D1ml, age) had a significant impact on the occurrence of LF in the test cohort.Conclusion: Despite the obvious difficulties in generalizing predictive models for oncologic outcome and toxicity, this analysis shows that a carefully designed radiomics model for prediction of local lung fibrosis after SBRT of early stage lung cancer performs well across different institutions.


Introduction
Stereotactic body radiation therapy (SBRT) is an effective therapy for early-stage, node-negative, medically inoperable non-small cell lung cancer (NSCLC). Dose-fractionation schemes usually depend on tumor size and location and have been largely standardized by current guidelines [1][2][3][4]. However, after irradiation, about 10-15% of the tumors will recur locally and up to 50% of the patients will experience systemic disease progression despite PET-based staging before SBRT [5]. Also, 25-30% of the patients will develop radiation-induced lung injury (RILI) on follow-up chest imaging, but apart from an established dose-response relationship for local control [6], dosimetric and clinical factors have only shown limited capability in predicting these events [7][8][9][10][11][12][13][14].
The availability of open-source software solutions allows the extraction of standardized radiomic features and generation of complex, non-linear models which are able to account for complex interactions between features and have the potential to achieve high performance. In the present study, we applied state-of-the-art feature extraction and machine-learning algorithms in order to determine the extra value of imaging tumor biomarkers when used in addition to dosimetric and clinical factors for prediction of radiation-induced lung injury, local control, disease free survival and overall survival in a cohort of patients with NSCLC treated by robotic SBRT. The model was trained on data from one institution and tested on a cohort from a separate institution that treated patients based on similar inclusion criteria and fractionation schemes. This work extends an earlier single institution report [36].

Patients, Treatment and Follow-Up
Two cohorts of patients with stage I/IIa NSCLC (according to staging classi cation of the Union for International Cancer Control [UICC], 8 th edition) who underwent de nitive robotic SBRT were retrospectively analyzed. The rst cohort comprised 110 patients treated at the University Hospital of Cologne, Germany and was used for identi cation of clinical, dosimetric and image-derived parameters to predict local control (LC), overall survival (OS), disease free survival (DFS) and occurrence of local lung brosis (LF) as a manifestation of radiation-induced lung injury after SBRT. This cohort had already been analyzed in a previous study using a different radiomics approach [36] and served as the training data set. A second cohort of 71 patients was treated at the Radiosurgery Center Northern Germany, Guestrow, and was used as test set (in machine learning terminology) for the predictive power of the identi ed parameters in the training set.
In both cohorts, patients suffering from a peripheral T1/2 (UICC 8) NSCLC without lymph node metastases who were either medically inoperable or refused resection were treated solely by means of the Cyberknife R system (Accuray, Sunnyvale, USA) without concomitant therapy using a risk-adapted fractionation scheme (peripheral T1 tumors 3x13-18Gy, T1 tumors with broad contact to the chest wall and T2 tumors 5x10-11Gy, near-central or true central tumors 8x6-7.5Gy). The dose was calculated using the MonteCarlo dose calculation algorithm and prescription and reporting was done according to international recommendations [13; 37-40]. The cohorts also contained 12(8) patients with local stage T1/2 tumors who had been successfully treated for oligo-metastatic disease, and who were free from tumor activity besides the primary tumor. Patient characteristics and treatment parameters are shown in Tab.1. All patients had a planning CT which was used for both treatment planning and radiomics image analysis (Tab. 2).
Clinical and radiological follow-up including chest CT scans was scheduled at 3 and 6 months after radiotherapy and every 6 months thereafter. A local recurrence was assumed if the irradiated lesion showed a solid core that increased by at least 25% compared to the last follow-up and exhibited further growth. Every occurrence of diffuse or patchy consolidation, diffuse or patchy ground glass opacity or modi ed or mass like consolidation in the lung tissue adjacent to the tumor was regarded as radiation induced lung injury and termed local brosis [41; 42]. Lung tissue changes smaller than the original tumor, scar-like patterns distant to the tumor and lung toxicities without clear spatial or temporal relation to radiotherapy (early acute pneumonia, late acute pneumonia, pneumonitis spatially not correlated to the PTV) were not considered. Representative chest CT images are shown in Fig. 4. In cases where a growing lesion could not be differentiated from local brosis, a FDG-PET-CT scan or a biopsy was performed in order to con rm or reject the diagnosis of a local recurrence.
Image processing and feature extraction Image processing was performed using Python 3.6.7 (Python Software Foundation, Beaverton, Oregon, USA). The original DICOM data containing manual delineations of the gross tumor volume (GTV) and anatomical image data were restored from the Cyberknife R archive and subsequently used to extract the target volumes for radiomic analysis.
For all further image processing, the software package pyradiomics 2.0.1 [43] was used that allows the extraction of standardized features which were de ned by the IBSI (Image Biomarker Standardization Initiative) [44]. Preprocessing included resampling to isotropic voxel of 1 mm 3 and removal of all pixels with Houns eld units (HU) below (-400) HU and above 1000 HU from the volume which were assumed to represent normal lung and bony tissue unintentionally included in the GTV. Radiomic features were calculated based on the original image and after wavelet ltering, yielding eight additional image types based on the application of wavelet-based high-pass or low-pass lters to each of the three dimensions. In addition to 14 features descriptive of the target's shape, 93 features were calculated for each of the nine image types, resulting in a total of 851 radiomic features.

Model development and statistical analysis
All model development was performed on the training cohort and the model parameters were optimized using crossvalidation schemes. First, the primary set of radiomics features was reduced by identifying and removing linearly correlated features with a Pearson correlation coe cient > 0.95. Out of the 851 extracted features, 564 were found to be highly linearly correlated and removed from the analysis. The remaining 287 features were then used to develop predictive models for each of the four endpoints: LC, OS, DFS and occurrence of local lung brosis (LF) after SBRT. In order to maximize generalizability and allow for complex non-linear relationships between feature values and treatment outcome, a gradient-boosted ensemble of regression trees was chosen as model. The algorithm is implemented in scikit-survival package for Python [45] and learns to predict the individual (log) hazard ratios from a combination of the radiomics features similar to a linear predictor of a Cox proportional hazards model [46]. Within the training set, the parameters of the model were subjected to a grid search with 5-fold cross-validation that regularized the depth of the regression trees (7), the learning rate (0.01) and the subset of features used for the next iteration (12 out of 287). The nal radiomics model was then evaluated by means of the concordance index and by using the log hazard ratios as predictive factors in the training and test sets.
In addition to the radiomics features, the following continuous clinical and dosimetric variables were analyzed in univariate Cox regression models with respect to their potential impact on any of the endpoints: GTV (gross tumor volume), PTV (planning target volume), GTV-D max (maximal dose in GTV), GTV-D mean (mean dose in GTV), GTV-D 95% (dose achieved in 95% of the GTV), PTV-D 95% (dose achieved in 95% of the PTV), Lung-D 1ml , Lung-D 10ml , Lung-D 50ml , Lung-D 100ml , tumor diameter, age and Charlson Comorbidity Score. Categorical clinical and treatment related factors were investigated using the Kaplan-Meier method and survival estimates were compared using two-sided log rank tests. These included: gender, T-Stage (T1 vs. T2), histology (squamous cell/ adeno /other/ unknown) and ducial tracking (no/ yes). Finally, all radiomics, clinical and dosimetric factors with signi cant impact in univariate analysis were evaluated in a multivariate Cox regression model. All statistical analyses were performed with the software R (version 3.4.4; R Development Core Team) or SPSS (vs. 24, Armonk, NY, USA). A p-value of <0.05 was considered signi cant. The complete work ow is depicted in Figure 1.

Clinical outcome
The outcome in terms of the analyzed clinical endpoints did not differ signi cantly between the two cohorts (Fig. 2). Overall survival at 36 months amounted to 56% vs. 43%, p=0.065), median DFS was 25 months vs. 23 months, p=0.43 and local control rates at 36 months were 90% vs. 93%, p=0.197). In the training set, none of the clinical and dosimetric factors had a signi cant in uence on the endpoints. Local lung brosis developed in 33% vs. 35% of the patients (p=0.75), all events were observed within 36 months after irradiation. As shown in Tab. 3, three dosimetric factors (GTV mean , PTV-D 95% , Lung-D 1ml ) and the patient's age had a signi cant impact on the development of local lung brosis with an increase in hazard of approximately 6% per Gy and per year of age. The direction and approximate size of these effects were also found in the test set but failed to reach statistical signi cance.

Radiomics model
In the training set, the algorithm was able to t highly predictive models for all endpoints using the remaining 287 radiomic features, resulting in concordance indices of 0.997, 0.996, 0.996, 0.998 for OS, LC, development of lung brosis and DFS, respectively. Consequently, the radiomics predictor was able to distinguish low-risk from high risk patients (log hazard ratio < 0 vs. > 0) for all of the endpoints OS, DFS, LC and LF with a high accuracy (p < 0.001, Fig.  3). The radiomics predictor for LF retained its predictive value when analyzed together with GTV mean , PTV-D 95% , Lung-D 1ml and age in a multivariate Cox Regression model (p<0.001, Tab. 3).
In the test cohort, the radiomics predictor was also able to differentiate patients with a high risk for developing LF from those with a low risk (p=0.016, concordance index of 0.635, Fig. 3). It also kept a signi cant in uence in a multivariate Cox regression model with the above mentioned dosimetric factors and age included (p<0.028). In contrast, the radiomics model failed to predict any of OS, DFS and LC in the test cohort (Tab. 3).

Summary of ndings
In the present analysis, two cohorts of early-stage lung cancer patients treated with robotic stereotactic body radiotherapy at two different institutions were investigated. Although slightly different fractionation schedules were applied, oncologic outcome in terms of local tumor control, disease-free survival and overall survival were well comparable. Importantly, the frequency and time course of development of radiation-induced local lung injury was also similar in the two cohorts. Radiomics analysis based on a set of standardized features and state-of-the-art modelling in the training cohort resulted in a model for prediction of radiation-induced local lung injury that performed well also in the test cohort and kept its predictive value when analyzed in conjunction with dosimetric parameters. However, the predictive models for the endpoints of oncologic outcome (OS, DFS, local control) failed to generalize to the test cohort.

Prediction of local radiation-induced lung injury
To the best of our knowledge, this is the rst report that generated a predictive, general model for the development of local lung injury from the GTV after lung SBRT [36]. Radiation-induced local lung injury that nally develops into local lung brosis is a typical event after lung SBRT, although it remains asymptomatic in most cases. It is probably triggered by the release of in ammatory cytokines such as TGF-ß from the tumor which subsequently initiate an immunological response [47; 48]. At rst sight, it seems far from obvious how a texture pattern detectable by radiomics could predict for this event. However, an association between a pre-therapeutic radiomics feature (LoG standard deviation) with the TGF-ß signaling pathway has recently been observed, and in the same report, a radiomics score was correlated with the amount of tumor in ltration by T-lymphocytes [49]. The view that image features correlate with the presence of immune-competent cells in lung tumor tissue is also supported by the observation that lung tumors characterized by low CT intensity and high CT heterogeneity exhibited a high CD3 (T-lymphocyte) in ltration, suggestive of an activated immune state [50].
In the present report, the radiomics model kept its predictive value for the development of lung brosis in both cohorts even when analyzed in a comprehensive model including dosimetric factors of the target volumes and adjacent lung (GTV-D mean , GTV-D max , PTV-D 95% , Lung-D 1ml ) although the predictive ability of the dosimetric factors could not be reproduced in the test cohort. In a comparable approach that has been applied for prediction of radiation pneumonitis from features of the total lung tissue in lung cancer patients treated with intensity-modulated radiotherapy (IMRT), the radiomics features only slightly improved the predictive value of the model when added to clinical and dosimetric factors [47]. Interestingly, the inhomogeneous dose distribution usually generated by robotic radiosurgery and volumetric arc therapy has itself been analyzed with respect to dose distribution patterns ("dosiomics") which in turn have been found to predict the incidence of radiation pneumonitis [51]. Thus, a more comprehensive model of radiation-induced lung injury could probably be built from incorporating texture analysis of the tumor, a shell [52; 53] comprising the adjacent lung tissue and the dose distribution.

Prediction of local control, disease-free survival and overall survival
Although the two cohorts resembled each other in terms of oncologic outcome, the radiomics model did not generalize from the training to the test cohort with respect to these endpoints. A compilation of recent studies on the impact of radiomics features on oncologic outcome for lung cancer patients after SBRT is presented in Tab. 4. Most of the studies applied single institution cross-validation or validation by test sets from the same institution and were able to predict local tumor recurrence, regional/nodal recurrence, distant failures and overall survival with a moderate accuracy. Of note, one report failed to observe features predictive of local recurrence [24]. Only in a minority of series were the results validated in test sets from independent institutions. In a large study from the Cleveland Clinic (Ohio, USA), a convolutional neural network (CNN) was trained to predict local recurrence in a group of > 900 lung cancer patients treated by SBRT. The strati cation resulted in two groups with highly signi cant different risk for recurrence in both the training and test set [54]. Also in another study where a CNN was applied to both CT and PET images, a highly accurate classi cation of survival probability was achieved in an independent data set [21].

Limitations Of The Present Study
The present study, although based on the results of two independent cohorts, probably still lacks a su cient number of patients needed for an informative analysis of the interaction between dosimetric parameters and radiologic tissue characteristics for prediction of local events (recurrence, local lung brosis) after SBRT of NSCLC. Also, the classi cation of local lung injury and tumor control is purely image-based and remains somewhat ambiguous, as tissue specimens are rarely available following SBRT. Differences in therapeutic strategies for detecting and treating metastases may have prevented the creation of a general radiomics-based model for prediction of DFS and OS.

Conclusions
The present analysis provides evidence that radiomics analysis can, in principle, be used for prediction of local lung injury after SBRT of NSCLC in independent data sets and as such complements existing results on the successful prediction of other oncologic endpoints in this setting.   Work ow for generating and validating the developed models.

Figure 2
Survival curves for overall survival (OS), local control (LC), disease free survival (DFS) and occurrence of local lung brosis after SBRT for the training and testing cohort. No signi cant difference between the cohorts was measured for any endpoint.  Kaplan-Meier curves displaying performance of the radiomics model in the training and test cohorts when stratifying patients into low and high risk groups Representative chest CT images of patients who did not (upper row) or did (lower row) develop local lung injury induced by robotic stereotactic body radiation therapy of early-stage non-small cell lung cancer